Claude Code: Back to the Terminal

The buzz

You can't escape it on X these past few weeks: everyone's talking about Claude Code. Developers raving, sharing screenshots, setups crazier than the last. Claude Code is installing an LLM on your machine, plugging a brain into your kernel, if you will. You enter a directory, type claude, and off you go—you're no longer alone, you're literally collaborating with the computer.

It's as spectacular to experience as it is hard to put into words.

I'd been using Cursor for months. I hesitated for a moment. Go back to the terminal? Really?

Yes.

The Cursor era

I've written about Cursor before: a powerful tool that transformed how I code. Visual control, transparent code generation, IDE integration, source code indexing, rules, model selection—all the good stuff. But there were two fundamental problems: cost and difficulty handling complex tasks (context limits).

Using Opus (Claude's most capable model) through Cursor quickly became a financial drain. $20 in two days, easily. Cursor tends to get greedy fast. If you don't rein it in, it reads tons of files and your context quickly hits hundreds of thousands of tokens—sometimes millions. At that rate, a single request costing $1-2 isn't unusual... With Opus, it's roughly $10 per day on average. And Opus is precisely what I need for Anantys—a complex project that demands nuance and deep code understanding.

I made do: I'd set up drastic Cursor rules to limit its greed (no reading files outside scope without my permission) and only used Opus for planning my development work (its job was to write a markdown spec), then handed off to Gemini 3 for actual coding (cheaper and not bad at code, even if a weaker architect than Opus).

It worked. But I'd ended up paying around $60-70/month.

First contact

So there I was, early January. I install Claude Code. I enter the Anantys repo, type claude. First impression: yes it's the terminal, but paradoxically, it's very ergonomic. Well-designed, polished.

First thing: I switch to Opus with the /model command. It makes a huge difference. Staying on Sonnet is like asking a high schooler to take over your keyboard, while Opus would be a Ph.D. In short.

First hour: over quota. BUT—and this matters—I was already impressed. What struck me immediately was the approval system. Claude Code asks for confirmation before every sensitive action. I see exactly what it's about to do. I approve or reject. This transparency, this explicit control, is exactly what I was looking for.

We follow the workflow together. I'm not watching a black box—I'm at the controls.

The pricing game-changer

I switched to the Max plan at $90/month. And then—revelation.

I never hit quota again. Limits reset every few hours, every week. With constant usage of a Claude Code instance—and I really do use it all the time—I never go above 70% of my daily limit.

Compare that to Cursor with Opus always on? We're talking $200-300/month easily. With Claude Code: fixed, predictable cost, and peace of mind. No more mental math wondering if I can afford a long session.

The features that make the difference

/init and CLAUDE.md

The /init command generates a CLAUDE.md file at the project root. It's the project "bible" for Claude. It reads conventions, architecture, key commands. Result: contextualized responses from the first prompt, without re-explaining everything each session.

Skills

In .claude/commands/, you can create "skills"—simple Markdown files containing prompts. The pattern is disarmingly simple: one file, one instruction, and Claude knows exactly what to do.

Concrete example: I created /anantys-review <scope> for code audits. A simple prompt that triggers a complete analysis of the requested scope. Simple + powerful = exactly right.

The productivity leap

Two weeks of intensive use. The verdict?

A massive leap forward.

With Cursor, I'd regularly go down false tracks—once or twice a week, entire sessions correcting poorly oriented code. With Claude Code? Near zero. The gaps are minor, not major rewrites. The code quality is remarkable and—critically—consistent.

It's not just "better"—it's a different regime entirely.

The new setup

I unsubscribed from Cursor.

Back to plain VSCode + the Claude Code extension. Best of both worlds: an editor I know inside out, and a command-line AI assistant that hides nothing from me.

Going further: Spec-Kit

A recent discovery for heavy projects: Spec-Kit. The idea: structure your specifications so Claude can consume them efficiently over long sessions.

I tested it on a real challenge: bank synchronization for Anantys. A complex feature—OAuth, webhooks, mapping to the local DB to link bank accounts to investment strategies. The kind of thing that normally takes several weeks.

With Spec-Kit + Claude Code: one day. I started in the morning, by 6pm I was testing with the provider's sandbox. It worked.

Plot twist: the provider's pricing wasn't what I expected (minimum license out of budget). A bit of a blow. But no reason to give up—multi-provider support was needed anyway (EU providers only support EU banks, and vice versa for the US).

New EPIC with Spec-Kit: refactor all the code into abstract classes to support any provider. 48 hours later: done, tested, functional. Implementing a new provider now comes down to coding its specific business class and converting its payloads to standardized Anantys objects.

The provider's sales rep was surprised my implementation was already working. They're used to assigning a "sales engineer" to help their clients...

2026: the reality check

Switching to Claude Code reminds me exactly of what I felt when I moved from hand-coding to Cursor. The same gap. The same feeling of stepping up a level.

Except this time, it's even stronger. I feel like I'm working with someone extremely qualified. Not a tool, not an assistant—a collaborator. One who understands context, proposes relevant solutions, and produces code I'm not ashamed to merge.

The numbers speak for themselves: an EPIC that would have taken me two to three weeks wraps up in one or two days. And I'm not sacrificing quality—quite the opposite. Opus's systematic reviews at the end of each session give me a safety net I never had before.

This isn't "coding with help" anymore. It's a radical transformation of what it means to be a developer. The world has changed. Completely.